App update kills all terminal sessions — no session persistence across Sparkle restart

[austinywang]

Summary

When cmux updates via Sparkle (auto-update or manual "Install Update"), the app quits and relaunches. All terminal sessions are destroyed in the process — running shells, background processes, SSH connections, and AI agents (Claude, Codex) are all killed. The update behaves like fully quitting the app and opening it fresh.

This is a major disruption for power users who may have dozens of workspaces with long-running processes. A single update wipes out all active work.

Expected behavior

Terminal sessions should survive app updates. Options include:

1. Session persistence across restart — serialize pty file descriptors and restore them after relaunch (like tmux/screen session detach/reattach)
2. Process preservation via fd passing — hand off pty FDs to a helper daemon before quit, then reclaim them on relaunch
3. Sparkle in-place update without restart — update the binary on disk but keep the running process alive (only restart on next manual quit)
4. At minimum: warn before updating — if sessions are active, show a confirmation dialog listing what will be lost ("12 workspaces with active processes will be terminated")

Current behavior

• Sparkle triggers app quit → relaunch
• All ptys are closed, sending SIGHUP to all child processes
• Shells, SSH sessions, running commands, and AI agents are all killed
• On relaunch, workspaces are recreated but with fresh empty shells
• No warning is shown about active sessions being destroyed

Suggested approach

The most robust solution is a session daemon (similar to how VS Code's terminal works, or tmux's server):

1. A lightweight cmuxd helper process owns the pty file descriptors
2. The GUI app connects to cmuxd to display terminal output
3. On update/restart, cmuxd keeps running — ptys stay alive
4. On relaunch, the GUI reconnects to cmuxd and reattaches to existing sessions

A simpler intermediate fix would be to at least warn the user before an update kills active sessions, and let them defer the update.

Impact

• Every update kills all running agents and terminal sessions
• Users lose unsaved work, SSH connections, long-running builds
• Discourages updating, leading to users running stale versions

[tiagoefreitas]

This PR #510 does this without needing a custom daemon, and is a more standard way, zmx is the best session persistence system based on libghostty-vt.

Using cmuxd to own the tty means its not compatible with other clients. It also benefits from updates to zmx.
Actually I would use zmosh that includes zmx for local tty but also supports attaching to remote zmx ttys (so its a stepping stone for remote cmux persistence too).

It allows multiple clients on the same session (so the same tty can be shown on multiple cmux tabs or remotely and locally at the same time) and native scrollback.

I am working on a proposal to make session persistent better with multiple client platforms (desktop,iphone,ipad) with a client-agnostic spec (and multiplexer agnostic, so cmux could work with tmux too or others), but I have little time now to finish it, so PR 510 is a good base for this (after fixing the issues the review bot mentioned).

[juanatsap]

This hit me hard too — a crash wiped an hour of carefully arranged workspaces. Splits, commands, everything. Gone.

Great to see progress on native session restore (#2544, #2545) and the zmx approach in PR #510. While those land, I built cmux-resurrect as an external tool that works today (the command is crex for short).

With crex save, you can snapshot your entire layout — splits, working directories, pinned tabs, startup commands. And with crex restore, bring it all back exactly as it was. It also supports auto-save via launchd, so you're covered even when Sparkle updates hit without warning.

There's quite more to it — like Workspace Blueprints (your layout as human-readable Markdown you can version in git), dry-run mode, and reusable templates — but save/restore is the main thing relevant to this issue.

It won't prevent the kill, but it guarantees recovery in seconds — which is something.

Thanks for building such amazing product!